2.1.

System Modules and Spectral Subsystems

The ODS design consists of three main components: a console, an optical probe head, and a single-use sheath (Fig. 1 ). The console contains all of the light sources, coupling optics, spectrometer, and electronics, and includes a calibration port for probe calibration prior to each patient scan. The optical probe is interfaced to the console by way of an articulating arm that supports the electrical and optical cables. A single-use sheath snaps onto the optical probe and protects the patient and the optics against cross-contamination. In use, the optical probe is positioned near the yolk of a vaginal speculum to view the cervix. The system is operated using custom software running on an embedded computer. The user interacts with the computer through a keyboard, a touch pad, and a dual-element foot pedal. The final system output is a video image of the cervix with a false-color overlay indicating areas of suspected high-grade disease.

Fig. 1

Photograph of the optical detection system showing some major features.

The spectral subsystems include the UV source, the broadband source, their respective coupling optics, the spectrometer and detector, the calibration system, the probe optics, the scanner assembly, the broadband illuminator, and the sheath. A detailed breakdown of the optical paths is shown in Fig. 2 . The transmit path for the UV source and the broadband sources differ. The UV source is used to induce tissue fluorescence, while the broadband source is used to measure diffuse reflectance. The spectrometer, detector, probe collection optics, and sheath window are common for fluorescence and diffuse reflectance measurements.

Fig. 2

Schematic of optical detection system.

The UV source is used to induce fluorescence from the cervix and other calibration targets. Light from a pulsed nitrogen laser (model 337201-99, Spectra-Physics, Mountain View, California; $200\text{-}\mathrm{\mu }\mathrm{J}$ nominal power output; $4.5\mathrm{ns}$ pulse width) is passed through a $337\text{-}\mathrm{nm}$ laser line filter, then shunted by a beamsplitter, sending 5% of the beam through a $60\text{-}\mathrm{deg}$ holographic diffuser into a UV-transmitting fiber directing laser energy onto one end of a 532-channel charge-coupled device (CCD) detector (S7031-0906, Hamamatsu Corporation, Bridgewater, New Jersey) (the CCD is split into spectral and energy-monitoring sections). The energy in this leg is used to monitor the laser energy and to correct for pulse-to-pulse energy fluctuations. Neutral density filters are used to control the energy on the detector in the energy-monitor beam and in the excitation beam delivered to tissue. The system is designed to deliver to tissue between 16 and $35\mathrm{\mu }\mathrm{J}$ per pulse of $337\text{-}\mathrm{nm}$ light.

The primary beam is coupled into a tapered 0.4- to 0.2-mm core, $0.22\text{-}\mathrm{NA}$ optical fiber that delivers the light to a set of collimating optics in the probe head. After reflecting off a dichroic beamsplitter and a 2-D galvo-driven stepping mirror (M3 Series, GSI Lumonics Incorporated), the beam is passed through a set of scan lenses designed to focus the beam to an in-focus diameter of $0.8\mathrm{mm}$ on the cervix. The scan lenses included a three lens system designed to have minimal chromatic and spherical aberrations between 330 and $720\mathrm{nm}$ and used low fluorescence I-line glass and antireflective coatings. The distance between the last lens and the focal plane is $130\mathrm{mm}$ . The scan mirror raster scans a $25\text{-}\mathrm{mm}\text{-}\mathrm{diam}$ circular region on the cervix by stepping back and forth through 499 individual, hexagonal-packed sites having a $1.1\text{-}\mathrm{mm}$ center-to-center distance between adjacent sites. 25 lines having varying number of measurement steps cover the $25\text{-}\mathrm{mm}\text{-}\mathrm{diam}$ circular region [see Fig. 3a ].

Fig. 3

Graphical representation of the spectral arbitration routines. (a) For this subject, only four measurement sites were lost due to glare being present in BB1 and BB2 (light gray circles), even though 173 sites were identified with glare or shadow in at least one channel. Combination includes sites with glare and shadow. (b) White light image of cervix showing the area scanned in (a) and includes a ring of glare.

Diffuse reflectance spectra are collected concurrently with the fluorescence data. Light from two 20-W xenon flashlamps (FX-1160, Perkin Elmer, Salem, Massachusetts) is coupled into separate optically integrating rods, and then each rod splits the light among three emergent fibers. The first fiber of each integrating rod delivers about 5% of the light to the CCD/spectrograph for energy monitoring. The rest of the light in each rod is split between two broadband fibers (PGR-FB2000, Toray Plastics, New York, $4.3\mathrm{m}$ in length) that deliver the light to the probe head. Thus, a total of four broadband transmit fibers (two from each of the rods) delivers light to four projection lenses that alternately flood illuminate the cervix from two angles (either from two upper sources or from two lower sources). The effective illumination area is $3.2\mathrm{cm}$ in diameter. The broadband emission spectrum is maximal at $487\mathrm{nm}$ and drops to about 10% of the peak at both 360 and $720\mathrm{nm}$ . Absorption occurring in the broadband transmit fibers and optics strongly attenuates light below $320\mathrm{nm}$ . The broadband energy incident onto tissue is set to approximately $300\mathrm{\mu }\mathrm{J}$ (range 290 to $310\mathrm{\mu }\mathrm{J}$ ).

The diffuse reflectance and fluorescence emission from cervical tissue are collected by the same scan lenses, passed through a dichroic beamsplitter and a $337\text{-}\mathrm{nm}$ excitation rejection filter, and then focused by the receive optics into a $0.2\text{-}\mathrm{mm}$ core, $0.22\text{-}\mathrm{NA}$ receive fiber. The receive fiber transmits the signal directly to the input of a broadband spectrograph (CP140-3301, Jobin Yvon OEM Group, Edison, New Jersey), which disperses the light onto the spectral portion of the CCD. Spectra are recorded between 360 and $720\mathrm{nm}$ at approximately $1\text{-}\mathrm{nm}$ resolution.

The ODS scans the cervix by collecting one fluorescence spectrum, two diffuse reflectance spectra (using the upper and lower light sources), and one background spectrum at each tissue measurement site. The background signal is measured within the same detector cycle. One detector cycle consists of five $3.3\text{-}\mathrm{ms}$ periods corresponding to a clear-detector period, a background measurement, the diffuse reflectance measurement with the upper lamps, the diffuse reflectance measurement with the lower lamps, and finally the fluorescence measurement. All detector measurements include a spectral region (CCD bins 11 to 401) and an energy-monitoring region (CCD bins 402 to 522). The 2-D detector dimension is 64 (imaging axis) by 532 (spectral axis) pixels. The imaging axis is summed into single bins.

2.2.

Calibration System

A full system calibration is performed routinely during system build and maintenance procedures using internal and external calibration targets and procedures. External targets include the use of a reference fluorescence dye cell, a calibrated nominal 60% diffuse reflectance target, and additional specialized video targets. The fluorescence intensity of the dye cell is a controlled reference that is used to correct across systems for the inherent differences in the collection efficiency of the different systems. The 60% target is used as an across-system reference for the diffuse reflectance measurement. The video targets consist of a split 10/20% diffuse reflectance target that is used to adjust the contrast and gain on the video camera, and a color target that is used to adjust the color balance.

Additional calibration sources include the use of argon and mercury pen lamps and a calibrated tungsten-halogen lamp. Internal targets and sources include a null target, a nominal 10% diffuse reflectance target, a small frequency doubled $\mathrm{Nd}:\mathrm{YAG}$ pen laser, a pulsed nitrogen laser, and two pulsed xenon lamps. The use of all targets and sources are described in more detail later.

In addition to routine full system calibration, additional run-time calibration procedures are made to check for changes that may occur between full system calibrations. Small changes are accounted for in the calibration procedures, while large changes prevent the user from proceeding to patient measurements. Run-time calibration is done just with the internal targets and sources. The number of calibration checks and targets are intentionally limited to improve the reliability of the system. Internal targets are housed within the system console (as part of a calibration port) and are automatically measured prior to each patient scan. The null target consists of a high optical density glass that returns negligible fluorescence and diffuse reflectance signal when illuminated by the probe. The nominal 10% diffuse reflectance target consists of a flat gray target modified to also contain a central $3.5\text{-}\mathrm{mm}\text{-}\mathrm{diam}$ plastic fluorescent plug and four peripheral phosphorescent plugs. The fluorescence and phosphorescent plugs are used for video and scan mirror alignment. In addition, the intensity of the fluorescence plug is monitored prior to each system use to ensure that the intensity is within predetermined reference values. The signal from the nominal 10% reflectance target is used in the diffuse reflectance calibration procedure (see next) and corrects for changes in the broadband source over time. The frequency double $\mathrm{Nd}:\mathrm{YAG}$ laser is used to account for temperature drifts observed in the spectrometer.

2.3.

Wavelength Calibration

The CCD spectral bins 11 to 401 are converted to wavelength by use of fundamental reference lines from mercury and argon pen lamps (models 6035 and 6033, Oriel Instruments, Stratford, Connecticut). Individual bins are converted to wavelengths using a fourth-order polynomial fit to the measured bin position of five mercury lines (365.01, 404.66, 435.84, 546.08, and $578.05\mathrm{nm}$ ) and two argon lines (696.54 and $738.40\mathrm{nm}$ ). In this way, wavelength calibration between 360 and $720\mathrm{nm}$ is determined to within $\pm 1\mathrm{nm}$ . Wavelength accuracy is also checked and corrected prior to each patient scan by use of the $532\text{-}\mathrm{nm}$ line from the focusing and alignment laser. If the measured value is within $3\mathrm{nm}$ of the expected value, the wavelength axis of the data is shifted to correspond to the measured laser line; if the deviation exceeds $3\mathrm{nm}$ , the scan is aborted and recalibration would be required. Systematic run-time wavelength drifts of $\pm 1.5\mathrm{nm}$ from full system calibration were observed in a small number of spectrometers (2 of 16) due to $10°\mathrm{C}$ temperature differences that occurred, depending on how long the system was turned on. Manufacturing changes to the grating mounts alleviated this dependency.

2.4.

Fluorescence Signal Calibration

The fluorescence signal from tissue is calibrated using a multistep procedure requiring 1. calibration data recorded at the time of full system calibration, and 2. calibration data recorded just prior to tissue measurements. The combined calibration elements are shown in Eq. 1 for each interrogation location $i$ , and at wavelength $\lambda$ .

The spectral sensitivity of the system is found by dividing the calibrated lamp signal $L$ by the calibration spectral irradiance values $W$ at each wavelength. The spectral sensitivity is divided by $\lambda$ to convert from watts to number of photons, then normalized to unity at ${\lambda }_n=500\mathrm{nm}$ . Planck’s radiation law for a blackbody emitter (at reference temperature $T$ ) is used to interpolate the spectral irradiance values measured at discrete wavelengths to the array of wavelengths at each detector bin.

The fluorescence signals from all systems are calibrated relative to a fluorescence dye standard ( $1.2\times {10}^{-5}\mathrm{M}$ Coumarin 515 in ethylene glycol). The peak of the dye fluorescence $\left({\lambda }_{\max }\right)$ is observed around $490\mathrm{nm}$ when dissolved in ethylene glycol. The term $D_c$ corresponds to the peak fluorescence intensity of the dye at $337\text{-}\mathrm{nm}$ excitation divided by the reference peak of the water $\mathrm{O}\text{-}\mathrm{H}$ stretch Raman vibration $(\sim 3350{\mathrm{cm}}^{-1})$ , which controls for calibration across dye lots. The value 176 corresponds to the ratio of the dye fluorescence peak to water Raman peak for the progenitor solution, which yielded an average of $879\mathrm{counts}∕\left(\mathrm{\mu }\mathrm{J}\text{-}\mathrm{nm}\right)$ measured on the first ten systems. Following calibration, fluorescence emission is reported in absolute units of $\mathrm{counts}∕\mathrm{nm}\text{-}\mathrm{\mu }\mathrm{J}$ .

2.5.

Diffuse Reflectance Signal Calibration

The reflectance signal from tissue is also calibrated using a multistep procedure requiring 1. calibration data recorded at the time of full system calibration, and 2. calibration data recorded just prior to tissue measurements. The combined calibration elements are shown in Eq. 2 for each interrogation location $i$ , and at wavelength $\lambda$ .

The 60% diffuse reflectance target is measured during full system calibration procedures. The term $C$ corresponds to a 10% diffuse reflectance target housed in the calibration port and is measured prior to patient measurements (first term) and the 60% target (second term) that is measured at the most recent full system calibration. The 10% target has four phosphorescent spots that are used for aligning the scan mirror relative to a defined optical axis, and a central fluorescence spot that is used as a calibration check. As a result, only a portion of this target is available for diffuse reflectance measurements precluding its use as the calibrated reflectance standard. The bracket $⟨\dots {⟩}_k$ implies an average of the 10% diffuse reflectance target over the 499 measurement sites, minus the sites covered by the phosphorescence and fluorescence targets. The mean 10% diffuse reflectance target is measured before each patient measurement and allows for corrections for changes in the light sources since the most recent maintenance procedure. As before, the diffuse reflectance signals are corrected for electronic background and external stray light $B$ , and internal stray light as measured by the null-target signal $N$ .

Following calibration, fluorescence and diffuse reflectance spectra are smoothed using a five-point median filter¹⁰ followed by a cubic 27-point Savitzky-Golay filter.^{11, 12} Following smoothing, the fluorescence and diffuse reflectance spectra are linearly interpolated to $1\text{-}\mathrm{nm}$ bins between 360 and $720\mathrm{nm}$ .

2.6.

Spectral Arbitration

At each scan interrogation location, one fluorescence spectrum and two diffuse reflectance spectra (illuminated with the upper and lower light sources) are available following spectral calibration. These reflectance spectra are occasionally subject to specular reflections (glare) or vignetting from the speculum (shadows). The reflectance spectra are evaluated for the presence of these effects by comparing differences in shape between spectra recorded with the upper and lower light sources. Specular reflection denoted as mirror-like reflection carries just the illumination spectral response absent tissue influence. As a result glare spectra are more intense and near constant lacking hemoglobin absorption bands, making them easy to distinguish. Diffuse reflectance spectra recorded when a portion of the illumination source is blocked by the speculum blade or vaginal wall tends to be lower when compared to the spectrum recorded with the other illumination source.

Glare in the diffuse reflectance spectra taken with each pair of light sources is identified using the intensities of the reflectance measurements in the 370- to $710\text{-}\mathrm{nm}$ spectral range. These two sets of measurements are also compared to evaluate the consistency of the measurements, with anomalous values being indicative of unwanted glare artifact or optical vignetting of one or both of the signals. Diffuse reflectance spectra determined to be affected by either specular reflection or vignetting are rejected. In cases where both spectra are free of such effects, the two spectra are averaged into a single spectrum. Additional treatment for glare missed by spectral arbitration is accomplished with an image processing module.

Fluorescence and reflectance spectra may also have abnormally low signals due to the measurement being made on blood, the os, or a speculum blade. Fluorescence signals at $479\mathrm{nm}$ and reflectance signals at $499\mathrm{nm}$ less than predetermined reference values are used to identify spectra having abnormally low signals. Tissue measurement sites having insufficient signal in either the fluorescence spectrum or both diffuse reflectance spectra are excluded from further consideration. Thus, combining the specular reflection, vignetting, and low signal metrics, the spectral arbitration process produces either no spectra, one fluorescence spectrum only, or one fluorescence and one diffuse reflectance spectrum from each measurement site.¹³

2.7.

Video Camera and Alignment and Focusing Systems

Prior to fluorescence and reflectance measurements, a color video camera (GP-KS462HM, Panasonic, Secaucus, New Jersey) captures an image of the cervix for analysis. Broadband light from both flashlamps is used to capture this image. The central optical axis of the video and the spectral channels are equivalent, as the two share most of the probe collection optics. Live video from the camera is also used to focus the ODS onto the cervix using a pattern generated by the alignment and focus laser source. An image of this pattern is captured to verify proper focusing prior to patient scans. Video images of the cervix are also captured once every second during scanning to detect movement of the cervix. Image capture is accomplished with a standard framegrabber board. Video images are captured by stepping the galvonometrically controlled scan mirror to the video position, which directs a view of the cervix toward the video camera collection optics (see Fig 2).

The alignment and focusing system is composed of a $10\text{-}\mathrm{mW}$ doubled $\mathrm{Nd}:\mathrm{YAG}$ laser source (model LCM-T-11ccs, Power Technology Incorporated, Little Rock, Arkansas), four fibers that are bundled together at the input end and have projection optics mounted to their output ends, and optics to couple the laser output into the fiber bundle. The laser output to tissue is designed to be less than $400\mathrm{\mu }\mathrm{W}$ at $532\mathrm{nm}$ .

Probe focus is determined using live video and the four projected green-laser spots arranged at the vertices of a rectangle centered on the cervix. These spots are generated by four illuminators positioned around the probe head at a relatively large angle with respect to the optical axis of the system. The spots move closer to the center of the image as the probe is pushed forward, and away from the center as the probe is pulled back. The probe position is adjusted until the four targeting spots are within four fixed circular targets overlaid on the video image of the cervix. Focus is validated by measuring the position of the spots and comparing them to a pre-established calibration curve that translates distance to probe location.

2.8.

Motion Tracking

Detection of translational motion of the cervix is accomplished with the aid of video images taken once every second during the course of the optical scan. The last one-second segment scan data is reacquired if the distance between the new and last image is more than $0.55\mathrm{mm}$ . The entire cervix is rescanned if the collective motion over the entire scan is greater than $2.5\mathrm{mm}$ .

The motion-tracking algorithm is based on the cross-correlation between the gradients of two consecutive images in the scan sequence.¹⁴ The cross-correlation $c(k,l)$ from the gradient images $I_1$ and $I_2$ is given by

3.1.

Patient Scan Data

ODS use on 2271 patients enrolled in evaluation studies was analyzed to assess clinical functionality and reliability of the system. Performance factors included the number of scans required for a result, the effectiveness of a single application of acetic acid, and the total optical exposure in use. Spectral data were evaluated to assess optical performance in clinical patients, including minimum spectral bandwidths required and signal-to-noise performance.

Repeat ODS scans may be necessary as a result of excessive patient motion, failure of the system to report a useful final result, or for an abort by the user for any reason. Of the 2271 patients studied, 96.7% (2197) required only a single full scan, while another 2.9% (65) had two scans, and 0.4% (9) had three or more scans. The reason for rescans was predominantly excessive motion, with the exception of seven cases that were for other reasons. Rescans include those due to partial rescans (due to motion $>0.55\mathrm{mm}$ or an optical power level outside of nominal values) and full rescans due to grossly excessive motion or a manual user abort.

The system prompts the user to apply acetic acid to the cervix and allows a scan to commence if it starts within 30 to $120\mathrm{s}$ after the application. Of the 2271 patients evaluated, 92.1% (2092) required only a single acetic acid application. The most common reasons for reapplication of acetic acid were due to difficulties focusing and aligning the probe prior to scanning. The elapsed time between when acetic acid was applied and the start of a scan was $73.5\mathrm{s}$ on average with a range between 31 and $123\mathrm{s}$ . Overall, the mean scan time was $11.9\mathrm{s}$ , and was less than $15\mathrm{s}$ in 99% of all cases. After a scan was completed, it took approximately $15\mathrm{s}$ for a final display result to be shown. Therefore, the total time from application of acetic acid to the display of a final result was typically about $2\min$ .

3.2.

Tissue Optical Exposures

The optical energy of the system was analyzed and found to be significantly below recognized standards for safe tissue exposure. Optical energy can come from one of three sources: the UV nitrogen laser, the frequency doubled $\mathrm{Nd}:\mathrm{YAG}$ alignment laser (green), and the broadband xenon lamps. The spectral irradiance of the xenon broadband sources was measured with a calibrated spectroradiometer. The output was found to be relatively flat over the specified 360- to $720\text{-}\mathrm{nm}$ spectral window, dropping rapidly below $360\mathrm{nm}$ to negligible levels at $320\mathrm{nm}$ and below. The system also has a built-in timer that does not allow the broadband sources to be on for more than $10\min$ per procedure, thereby strictly limiting optical exposure.

The system does not deliver more than $10\mathrm{UV}$ laser pulses to any one area of the cervix, again for safety reasons. A value of $35\mathrm{\mu }\mathrm{J}$ (mean plus six standard deviations) is used as the worse case single pulse exposure. Using these data and following American Conference of Governmental and Industrial Hygienists (ACGIH) guidelines,¹⁵ the maximum UV exposure to tissue for the system was determined to be 70 times less than the threshold limiting value (Table 1 ), even when considering the additional UV exposure delivered to tissue from a standard colposcope (OPMI-1FC, Carl Zeiss). The UV output of the OPMI-1FC was significantly lower than previously reported for Leisegang colposcopes.¹⁶ The OPMI-1FC transmits light to the cervix through a fiber optic bundle that effectively filters out light below $370\mathrm{nm}$ . The Leisegang colposcope directly illuminates the cervix using small diameter, thin lenses and delivers more UV light to the cervix. Optical testing determined that the ODS as used is a class 1 laser product, meaning that the laser output from the system from both the UV and green lasers is safe under any normal use circumstance.

Table 1

UV optical exposure for optical detection system using worst case exposure conditions and comparison to the threshold limiting value (TLV). Analysis is based on the American Conference of Governmental and Industrial Hygienist (ACGIH) guidelines.15 The safety ratio is calculated based on a TLV of 3000μJ∕cm2 . The spectral radiant UV exposure of UV laser result at 32,000 was measured at 337nm .

3.3.

Aggregate Spectral Files

Aggregate diffuse reflectance and fluorescence spectra of relatively pure cervical histologic classes, and other colposcopically identified groups, were compiled as a result of a pilot study.⁹ Five tissue classes were histologically diagnosed (normal squamous epithelium, normal columnar, metaplasia, CIN 1, and CIN 2,3), and one class was formed by colposcopic impression (normal squamous by impression). Seven additional classes representing obstructions were also identified. The number of patients and spectra in each class are listed in Table 2 . These data were used to evaluate the functional performance of the optical system.

Table 2

Aggregate groups and results of spectral arbitration. Column three shows reflectance spectra affected by glare G , or vignetting V , when illuminated with the upper 1 or lower 2 light sources. Column four shows measurement sites lost due to low fluorescence or reflectance signals (LS) , glare from both upper and lower light sources (GG) , glare from the upper and vignetting from the lower light sources (GV) , vignetting from the upper and glare from the lower light sources (VG) , and vignetting from both light sources (VV) .

3.4.

Spectral Arbitration

The results of spectral arbitration on the 13 aggregate groups are also given in Table 2. As designed, the number of spectra lost due to arbitration is relatively low for the tissue aggregate groups and relatively high for the obstruction groups. The low signal metric captures 2469 of the 3212 sites (77%) recorded on the blades of a metal speculum. A good fraction of blood (72%) and fluids (26%) are filtered by the low signal metric. In the glare group, all 2095 sites have glare in the reflectance data when recorded with the upper or lower light sources, yet only 302 sites (14%) are lost due to glare occurring simultaneously in both spectra. Closer to 50% of the sites would have been lost due to glare had a single light source been used. Cervical edges typically denote the transition between the lower speculum blade and cervical tissue, and, as expected, this group has a higher incidence of vignetting in the reflectance spectra taken with the lower light source. A majority of these sites, however, are preserved due to measurements taken with the upper light source. In a typical scan, both diffuse reflectance spectra from a given site are excluded in less than 5% of the 499 sites measured per cervix.

The results of spectral arbitration can be seen visually in Fig. 3, where the output of arbitration is depicted graphically alongside the color image of the cervix. In this extreme case, 173 of 499 sites were affected by glare or shadowing of the illumination sources, yet data were only lost from four sites in total (four sites had glare in both channels, zero sites had shadow in both channels, and zero sites had combined glare and shadow).

3.5.

Signal and Signal to Noise

Following calibration, fluorescence emission is reported in absolute units of $\mathrm{counts}∕\mathrm{nm}\text{-}\mathrm{\mu }\mathrm{J}$ and diffuse reflectance is reported as a percentage relative to light reflected from a calibrated diffuse reflectance target. The pilot study optical detection systems had an average UV laser energy at tissue of $22\pm 3\mathrm{\mu }\mathrm{J}∕\mathrm{pulse}$ , an average photoelectron conversion factor of $10.2\pm 0.4$ photoelectrons per count, and a typical quantum efficiency of 0.755 at $500\mathrm{nm}$ . Combining these values yields a factor of 297 that can be used to approximate counts $\mathrm{detected}∕\left(\mathrm{\mu }\mathrm{J}\text{-}\mathrm{nm}\right)$ to photons $\mathrm{detected}∕\mathrm{nm}$ .

The inherent spectral resolution of the spectrograph, including its $200\text{-}\mathrm{\mu }\mathrm{m}$ input fiber, is $8.3\mathrm{nm}$ . In subsequent data processing, a Savitzky-Golay smoothing filter produces an effective spectral resolution of $15.6\mathrm{nm}$ as determined by measuring the full-width at half maximum (FWHM) of a narrow reference band (data not shown). The bandwidth of the digital filter was selected based on tissue spectral properties and on the output from the classification algorithm. Typical fluorescence and diffuse reflectance spectra from clinical measurements subsequently identified as CIN 2,3 and normal squamous cervical tissue are shown with and without smoothing in Fig. 4 . It is observed that the high frequency noise seen in the unsmoothed data is effectively removed by the digital filters without introducing any distortion to the spectral line shapes. The smoothing routine occurs as one of the last steps in the calibration procedure and is done with the final calibrated spectra. No filtering is performed on the individual calibration spectra, raw tissue, or background spectra. The performance of the classification algorithm did not change for data measured with this system for bandwidths up to 27 points (data not shown).

Fig. 4

Typical $\left(a\right)$ diffuse reflectance and $\left(b\right)$ fluorescence spectra of cervical tissue following application of acetic acid. Individual points represent the data before Savitzky-Golay smoothing.

The signal-to-noise ratios for the spectra were generally very high. Signal-to-noise ratios exceeding 100 were typical over most of the wavelength range for both diffuse reflectance and fluorescence spectra after smoothing. For typical signals from tissue, the signal-to-noise ratios for diffuse reflectance spectra were generally higher than the fluorescence spectra due to the higher signals collected in the reflectance channel.

3.6.

System Validation

The fluorescence and diffuse reflectance spectra were recorded from known standards in ten systems to assess the variability among systems and the deviation from reference values. Ten percent diffuse reflectance targets were used to validate the diffuse reflectance channel [Fig. 5a ]. The mean(SD) diffuse reflectance signal measured at $500\mathrm{nm}$ for the ten systems was 9.5%(0.4%) corresponding to a 3.8% variation among systems and a $-5\%$ relative deviation from the expected 10% value. The fluorescence dye cells were used to validate the fluorescence channel [Fig. 5b]. The mean(SD) fluorescence signal was $871\left(37\right)\mathrm{counts}∕\mathrm{\mu }\mathrm{J}\text{-}\mathrm{nm}$ , corresponding to a variation among systems of 4.2% and a relative deviation from the expected value of $-0.9\%$ .

Fig. 5

Standard $\left(a\right)$ diffuse reflectance target and $\left(b\right)$ fluorescence dye spectra as measured on ten calibrated pivotal study systems. The intersection of the dashed lines depicts the true value.

Wavelength validation was determined by monitoring the position of the Coumarin-515 peak position. The mean(SD) peak position was $489.9\left(0.7\right)\mathrm{nm}$ (range 489 to $491\mathrm{nm}$ ). In addition, the position of $1\text{-}\mathrm{nm}$ narrow spectral bands selected between 360 and $720\mathrm{nm}$ from a commercial spectrophotometer were measured with three systems, and all bands were found to be within $\pm 1\mathrm{nm}$ from the known output wavelength.